echodict/llama.cpp

1#include "models.h"2 3llm_build_exaone::llm_build_exaone(const llama_model & model, const llm_graph_params & params) :4    llm_graph_context(params) {5    const int64_t n_embd_head = hparams.n_embd_head_v();6 7    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());8    GGML_ASSERT(n_embd_head == n_rot);9 10    ggml_tensor * cur;11    ggml_tensor * inpL;12 13    inpL = build_inp_embd(model.tok_embd);14 15    // inp_pos - contains the positions16    ggml_tensor * inp_pos = build_inp_pos();17 18    auto * inp_attn = build_attn_inp_kv();19 20    ggml_tensor * inp_out_ids = build_inp_out_ids();21 22    for (int il = 0; il < n_layer; ++il) {23        ggml_tensor * inpSA = inpL;24 25        // norm26        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);27        cb(cur, "attn_norm", il);28 29        // self-attention30        {31            // rope freq factors for llama3; may return nullptr for llama2 and other models32            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);33 34            // compute Q and K and RoPE them35            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,36                    n_embd_head, n_head, n_head_kv, il);37 38            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,39                                 ext_factor, attn_factor, beta_fast, beta_slow);40 41            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,42                                 ext_factor, attn_factor, beta_fast, beta_slow);43 44            cb(Qcur, "Qcur", il);45            cb(Kcur, "Kcur", il);46            cb(Vcur, "Vcur", il);47 48            cur = build_attn(inp_attn,49                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,50                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);51        }52        if (il == n_layer - 1 && inp_out_ids) {53            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);54            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);55        }56        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);57        cb(ffn_inp, "ffn_inp", il);58 59        // feed-forward network60        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);61        cb(cur, "ffn_norm", il);62 63        cur = build_ffn(cur,64                model.layers[il].ffn_up, NULL, NULL,65                model.layers[il].ffn_gate, NULL, NULL,66                model.layers[il].ffn_down, NULL, NULL,67                NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);68        cb(cur, "ffn_out", il);69 70        cur = ggml_add(ctx0, cur, ffn_inp);71        cb(cur, "ffn_out", il);72 73        cur = build_cvec(cur, il);74        cb(cur, "l_out", il);75 76        // input for next layer77        inpL = cur;78    }79    cur = inpL;80 81    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);82 83    cb(cur, "result_norm", -1);84    res->t_embd = cur;85 86    // lm_head87    cur = build_lora_mm(model.output, cur);88 89    cb(cur, "result_output", -1);90    res->t_logits = cur;91 92    ggml_build_forward_expand(gf, cur);93}94